06-reference

nateherk 3 ways to deploy claude agents

2026-05-15·reference·source: Nate Herk (X long-form)·by Nate Herk
claude-code-deployment/loopclaude-routinesmodaltrigger-devclaude-agent-sdkagent-orchestrationharness-engineeringautonomous-loops

"I Tested 3 Ways to Deploy Claude Agents (Here's When to Use Each)" - @Nate Herk

Structural-bias note: author runs AI Automation Society (700k YouTube) and discusses Modal / Trigger.dev as cloud-deployment options. No explicit sponsor block, but the piece functions as a survey-with-rubric that flatters all three vendors as fit-for-purpose. Read as a framework, not a buying recommendation.

Why this is in the vault

Direct meta-architecture for the RDCO autonomous-agent stack. The piece names and rubric-orders four deployment shapes (Claude Code /loop, Claude Routines, Modal/Trigger.dev, Agent SDK) on two axes (where it runs, how agentic), and surfaces the May 13 2026 Anthropic announcement that subscription credits now flow to Agent SDK - the cost question that previously blocked Method 3 + agent for us. This is the first survey-style framing of our entire deployment surface I've seen written down by someone outside our context, and it maps cleanly onto what we're already running.

The core argument

Two axes determine deployment shape: (1) where the agent runs - your machine vs someone's cloud - and (2) how agentic it is - full autonomous reasoning loop vs deterministic script. From these axes, four methods:

Decision rubric:

Punchline: "Match the method to the job."

Mapping against Ray Data Co

RDCO IS Method 1, hardened. Our entire autonomous-loop stack (vault-health, process-inbox, check-board, process-newsletter watch, morning-prep, finance-pulse, self-review, improve, graph-reingest, curiosity, deep-research, sync-contacts, process-youtube watch - 13 active crons) runs as Claude Code /loop jobs inside tmux on the Mac Mini, with a LaunchAgent supervising and a daily 4am session restart. That's production-shaped Method 1. Nate's framing validates the architecture; it does NOT suggest a migration.

Two-loop /clear trick is a finer-grained context-hygiene option we don't currently use. Our context hygiene is coarse: tmux + daily 4am restart. Nate's pattern (one /loop fires the work, a second /loop fires /clear every 5 min) would give per-cycle context resets without restarting the session. Worth piloting on one cron - probably /process-newsletter watch since it ingests the largest artifacts - to measure context-drift improvement before the next compaction. Queued for /improve review when bandwidth opens, NOT a today-decision.

Method 2 (Claude Routines) is not for us. 15/day cap on Max would not fit our 13-loop stack, and we'd lose local filesystem access (vault writes, audit scripts, qmd index, /tmp caches, attachment reads) that ~half our skills depend on. Don't migrate.

Method 3 candidates exist. Several cron loops don't need agent reasoning to fire - /sync-contacts (Gmail + Calendar query into ledger write), /graph-reingest (Python script + qmd), and the deterministic audit-script post-condition checks could migrate to Modal or Trigger.dev as straight scripts and free up Claude Code session slots. Future-architecture question, NOT today. Queued for /improve review when bandwidth opens.

Agent SDK is the long-term Method 3 + agent shape. The May 13 announcement that subscription credits flow to Agent SDK is the cost unlock that previously blocked this path. Worth a deeper read when we hit the next "RDCO infrastructure agent runs without a Mac Mini" question. Not relevant until then.

Hooks - we use them lightly. Yesterday's IndyDevDan agentic-security piece flagged a global pre-tool-use bash hook for RDCO settings.json as a gap; that aligns directly with Nate's "deterministic glue on top of agentic flows" framing. Already queued.

Queue for /improve review (low-cost experiments, no founder decision needed today):

Related